Improving efficiency of general surgery department by dynamic surgical scheduling with emergency uncertainty

Surgical scheduling is a complex, stochastic, and dynamic task in hospital resource planning. Maximizing the use of limited surgical resources and improving the efficiency of general surgery departments remain key challenges in healthcare operations management. This study aims to achieve real-time optimization and intelligent decision-making for emergency surgical scheduling (ESS) using deep reinforcement learning (DRL). The goal of ESS is to minimize both the maximum and average completion times for all patients while maximizing the average utilization rates of operating rooms (ORs) and non-operating room (NOR) beds. The ESS process can be modeled as a multi-action reinforcement learning task, formulated as a multiple Markov decision process (MDP). A graph isomorphism network (GIN) based multi-point graph network (MPGN) is used to process the graph representation, and a bi-proximal policy optimization (Bi-PPO) algorithm is developed to solve the MDP-based ESS problem. Four composite dispatching rules are designed to simultaneously select an unprocessed surgery and assign it to an available OR or NOR bed. At each rescheduling point, the Bi-PPO algorithm selects the most suitable dispatching rule by continuously training and optimizing the scheduling policy network. Numerical experiments across different surgical case scales are conducted to evaluate the effectiveness and generalization capability of the proposed Bi-PPO approach.